VA memcpy: HBM support

3 files changed, 63 insertions, 25 deletions

diff --git a/VA/baselines/cpu/app_baseline.c b/VA/baselines/cpu/app_baseline.c
index 84f1486..f5981ac 100644
--- a/VA/baselines/cpu/app_baseline.c
+++ b/VA/baselines/cpu/app_baseline.c
@@ -39,6 +39,7 @@ static T *B;
 static T *C;
 
 #if NUMA_MEMCPY
+int numa_node_local = -1;
 int numa_node_in_is_local = 0;
 static T *A_local;
 static T *B_local;
@@ -109,7 +110,7 @@ struct Params input_params(int argc, char **argv) {
 #endif
 
     int opt;
-    while((opt = getopt(argc, argv, "hi:w:e:x:t:a:b:c:")) >= 0) {
+    while((opt = getopt(argc, argv, "hi:w:e:x:t:a:b:c:C:")) >= 0) {
         switch(opt) {
         case 'h':
         usage();
@@ -123,11 +124,9 @@ struct Params input_params(int argc, char **argv) {
 #if NUMA
         case 'a': p.bitmask_in    = numa_parse_nodestring(optarg); break;
         case 'b': p.bitmask_out   = numa_parse_nodestring(optarg); break;
-#if NUMA_MEMCPY
-        case 'c': p.numa_node_cpu = atoi(optarg);
-                  p.bitmask_cpu   = numa_parse_nodestring(optarg); break;
-#else
         case 'c': p.numa_node_cpu = atoi(optarg); break;
+#if NUMA_MEMCPY
+        case 'C': p.bitmask_cpu   = numa_parse_nodestring(optarg); break;
 #endif // NUMA_MEMCPY
 #endif // NUMA
         default:
@@ -254,6 +253,16 @@ int main(int argc, char **argv) {
             A_local = A;
             B_local = B;
         }
+        mp_pages[0] = A_local;
+        if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
+            perror("move_pages(A_local)");
+        }
+        else if (mp_status[0] < 0) {
+            printf("move_pages error: %d", mp_status[0]);
+        }
+        else {
+            numa_node_local = mp_status[0];
+        }
         stop(&timer, 2);
 #endif
 
@@ -278,10 +287,10 @@ int main(int argc, char **argv) {
         if (rep >= p.n_warmup) {
 #if NUMA_MEMCPY
             printf("[::] VA-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%d"
-                " numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
+                " numa_node_in=%d numa_node_local=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
                 " | throughput_MBps=%f",
                 nr_threads, XSTR(T), input_size,
-                numa_node_in, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
+                numa_node_in, numa_node_local, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
                 input_size * 3 * sizeof(T) / timer.time[0]);
             printf(" throughput_MOpps=%f",
                 input_size / timer.time[0]);
diff --git a/VA/dimes-hetsim-hbm.sh b/VA/dimes-hetsim-hbm.sh
index a7c9ef4..7c0df73 100755
--- a/VA/dimes-hetsim-hbm.sh
+++ b/VA/dimes-hetsim-hbm.sh
@@ -1,7 +1,6 @@
 #!/bin/sh
 
 cd baselines/cpu
-make -B NUMA=1
 
 mkdir -p log/$(hostname)
 fn=log/$(hostname)/dimes-hetsim-hbm
@@ -10,22 +9,49 @@ fn=log/$(hostname)/dimes-hetsim-hbm
 
 (
 
-echo "single-node execution (1/2)" >&2
+make -B NUMA=1 NUMA_MEMCPY=1
+
+echo "CPU single-node operation with setup cost, cpu/out on same node (1/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
-	./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 5 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -C {ram_local} -t {nr_threads} -w 0 -e 40 \
 	::: nr_threads 1 2 4 8 12 16 \
-	::: cpu $(seq 0 7) \
-	::: ram $(seq 0 15) \
+	::: ram_in $(seq 0 15) \
+	:::        cpu $(seq 0 7) $(seq 0 7) \
+	:::+ ram_local $(seq 0 15) \
+	:::+   ram_out $(seq 0 15) \
 	::: input_size 167772160
 
-echo "multi-node execution (2/2)" >&2
+make -B NUMA=1
+
+echo "single-node execution, cpu/out on same node (2/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
-	./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 5 \
+	::: nr_threads 1 2 4 8 12 16 \
+	:::   ram_in $(seq 0 15) \
+	:::      cpu $(seq 0 7) $(seq 0 7) \
+	:::+ ram_out $(seq 0 15) \
+	::: input_size 167772160
+
+echo "single-node execution, in/out on same node (3/4)" >&2
+
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 5 \
+	::: nr_threads 1 2 4 8 12 16 \
+	::: cpu $(seq 0 7) \
+	:::   ram_in $(seq 0 15) \
+	:::+ ram_out $(seq 0 15) \
+	::: input_size 167772160
+
+echo "multi-node execution (4/4)" >&2
+
+parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \
 	::: nr_threads 32 48 64 96 128 \
 	::: cpu -1 \
-	::: ram $(seq 0 15) \
+	::: ram_in $(seq 0 15) \
+	::: ram_out $(seq 0 15) \
 	::: input_size 167772160
 
 ) >> ${fn}.txt
diff --git a/VA/dimes-hetsim-nmc.sh b/VA/dimes-hetsim-nmc.sh
index fa9e975..1e33c09 100755
--- a/VA/dimes-hetsim-nmc.sh
+++ b/VA/dimes-hetsim-nmc.sh
@@ -40,20 +40,22 @@ make -B NUMA=1
 
 (
 
-echo "CPU single-node operation (1/2)" >&2
+echo "CPU single-node operation (1/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
-	./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \
-	::: ram 0 1 \
-	::: cpu 0 1 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \
+	::: ram_in 0 1 \
+	:::      cpu 0 1 \
+	:::+ ram_out 0 1 \
 	::: nr_threads 1 2 4 8 12 16 \
 	::: input_size 167772160
 
-echo "CPU multi-node operation (1/2)" >&2
+echo "CPU multi-node operation (2/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
-	./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \
-	::: ram 0 1 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \
+	::: ram_in 0 1 \
+	::: ram_out 0 1 \
 	::: cpu -1 \
 	::: nr_threads 24 32 \
 	::: input_size 167772160
@@ -66,10 +68,11 @@ make -B NUMA=1 NUMA_MEMCPY=1
 
 echo "CPU single-node operation with setup cost" >&2
 
-parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
-	./va -i {input_size} -a {ram} -b {cpu} -c {cpu} -t {nr_threads} -w 0 -e 40 \
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
+	./va -i {input_size} -a {ram} -b {ram_out} -c {cpu} -C {cpu} -t {nr_threads} -w 0 -e 40 \
 	::: ram 0 1 \
-	::: cpu 0 1 \
+	:::      cpu 0 1 \
+	:::+ ram_out 0 1 \
 	::: nr_threads 1 2 4 8 12 16 \
 	::: input_size 167772160